A lightweight model of ShuffleNetV2-EH with higher accuracy of identification， lower complexity of computation， and suitable for identifying the types of feed raw material based on the lightweight convolutional neural network model ShuffleNetV2 to achieve rapid identification of warehousing feed raw materials and solve the difficulties in manually identifying the types of feed raw materials with similar crushing degree， color， and shape in currently processing and producing the combined feed raw materials. Firstly， the efficient channel attention（ECA） mechanism was introduced into the structure of ShuffleNetV2 network model， which adaptively adjusts channel weights based on input to enhance the ability of network model to percept important features in images of feed raw materials. Secondly， ReLU was replaced with HardSwish activation function to improve the recognition accuracy of the model without adding additional weights and parameters of bias. Finally， the structure of ShuffleNetV2 network model was adjusted to reduce the number of parameters and the complexity of computation in the model on the basis of ensuring the recognition accuracy of model. The results showed that the recognition accuracy of ShuffleNetV2-EH model on image sets from 8 types of feed raw materials was 99.13%， 1.38% higher than that of the original ShuffleNetV2 model. Its accuracy， recall， and F₁ score increased by 1.45%， 1.63%， and 1.62 %， respectively. The number of model parameters and floating-point operations decreased by 352 092 and 45.27×10⁶， compared to that of the original model. The overall performance was superior to classical convolutional neural network models including AlexNet， VggNet16， GoogLeNet， and ResNet18. It is indicated that the improved ShuffleNetV2 model well balances the complexity of computation and the recognition accuracy of the model， providing an algorithm foundation for online identification of feed raw materials in the warehousing process.